Research map / Large language models / Agents, code and math
Language Models and Logic Programs for Trustworthy Tax Reasoning
William Jurayj, Nils Holzenberger, Benjamin Van Durme
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add retrieved gold statutes and 5 gold exemplar parses (few-shot parsing) instead of zero-shot parsing (ablation) | GPT-4.1 · SARA v2 (100 tax calculation cases) | correct cases: 39 -> 87 | Improved |
| add retrieved gold statutes and exemplar parses instead of zero-shot parsing (ablation) | o3 · SARA v2 (100 tax calculation cases) | correct cases: 75 -> 81 | Mixed |
| use parsing to Prolog solver instead of direct calculation (zero-shot) | o3 · SARA v2 (100 tax calculation cases) | break-even price: $6,431.84 -> $47.43 | Improved |
| use parsing to Prolog solver instead of direct calculation (zero-shot) | DeepSeek-V3 · SARA v2 (100 tax calculation cases) | break-even price: $739.45 -> $2,099.13 | Hurt |
| use parsing instead of direct calculation (zero-shot) | small models (Qwen-32b, Llama-70b, R1 distills) · SARA v2 (100 tax calculation cases) | correct cases: Qwen-32b 13 -> 2; R1-70b 43 -> 2 | Hurt |
| add self-check with two independent solutions (Few-Shot + Few-Shot) (ablation) | GPT-4.1 · SARA v2 (100 tax calculation cases), gold statutes and exemplars | break-even price: $247.99 -> $40.08 | Improved |
| add self-check with two independent solutions (Direct + Direct) (ablation) | DeepSeek-R1 · SARA v2 (100 tax calculation cases), zero-shot (no gold statutes) | break-even price: $304.29 -> $94.20 | Improved |
| add self-check with two independent solutions (Few-Shot + Few-Shot) (ablation) | GPT-5 · SARA v2 (100 tax calculation cases), gold statutes and exemplars | break-even price: $15.78 -> $29.28 | Hurt |
| use reasoning-optimized model instead of chat model for few-shot parsing | DeepSeek-V3 vs DeepSeek-R1 · SARA v2 (100 tax calculation cases), gold statutes and exemplars | correct cases: 78 -> 40 | Hurt |
| use reasoning-optimized model instead of chat model for direct solving | DeepSeek-V3 vs DeepSeek-R1 · SARA v2 (100 tax calculation cases), zero-shot (no gold statutes) | correct cases: 22 -> 74 | Improved |
| longer case text | all models · SARA v2 (100 tax calculation cases) | failure rate | Hurt |
Closest papers by meaning
- The SEC Filings Dataset: Reconstructing U.S. Corporate and Financial Disclosures into Layout-Faithful and Token-Efficient Pretraining Data (NeurIPS 2026 · similarity 0.83)
- Neurosymbolic Learning for Inference-Time Argumentation (NeurIPS 2026 · similarity 0.83)
- Argumentative Large Language Models for Explainable and Contestable Claim Verification (AAAI 2025 · similarity 0.82)
- LexChain: Modeling Legal Reasoning Chains for Chinese Tort Case Analysis (AAAI 2026 · similarity 0.82)
- Are Language Models Efficient Reasoners? A Perspective from Logic Programming (NeurIPS 2025 · similarity 0.82)
- Detecting Bugs with Substantial Monetary Consequences by LLM and Rule-based Reasoning (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
